Editorial illustration for Quark Full W4A4 0.0161 VBench Gap Matches Wan2.2 Video Quality Drop on AMD GPUs
4-Bit Quantization Drops Video Quality on AMD GPUs
Quark Full W4A4 0.0161 VBench Gap Matches Wan2.2 Video Quality Drop on AMD GPUs
AMD's ROCm team ran a quantization experiment on Wan2.2-T2V-A14B, a large diffusion transformer used for text-to-video generation, and found that shrinking every layer to 4-bit weights and 4-bit activations produces a visible quality drop, even when the VBench score looks close to the full-precision BF16 baseline. Full W4A4 MXFP4 quantization scored 0.8520 on VBench overall. That's a small numerical gap from BF16, but the blog's authors say the gap shows up clearly in the generated video itself.
The fix they tested isn't quantizing everything the same way. Using AMD Quark, the team evaluated each transformer block individually and promoted only the blocks sensitive to precision loss from W4A4 to W4A8, leaving the rest on the cheaper 4-bit path. That calibration-driven mix pushed the score to 0.8648. A separate test quantizing the entire model to W4A8 using round-to-nearest (RTN) hit 0.8658, the highest quantized score measured in the study, suggesting FP8 activations applied model-wide can close most of the quality loss on their own.
All the W4A8 figures here come from quantize-dequantize emulation, not live hardware execution, so this is a quality study rather than a speed benchmark. Native W4A8 runs are supported starting in AMD Quark 0.13.
The safest reading is that full W4A8 and fixed-rotation/SmoothQuant-selected W4A8 are both near-BF16 quality, while the selected policy is the more practical current path because it uses W4A8 only where calibration indicates a benefit.
Why this matters
For anyone running video generation at scale on AMD Instinct hardware, this result is a useful sanity check rather than a victory lap. The AMD team tested how far they could push quantization on Wan2.2 before quality visibly degrades, and they found the floor: full W4A4 weights and activations produce a 0.0161 VBench gap that human reviewers actually noticed. That's the number worth remembering if you're tempted to squeeze every last bit out of your inference budget.
The more interesting finding is that W4A8, whether applied broadly or targeted through fixed-rotation and SmoothQuant selection, stays close to BF16 quality. That gives teams building on Wan2.2 or similar diffusion transformers a concrete reference point for where the precision/quality tradeoff actually breaks, instead of guessing. For founders evaluating AMD as a serving platform against Nvidia-centric stacks, this kind of granular benchmarking matters more than marketing claims about "efficient" quantization.
It won't settle the hardware debate, but it narrows the uncertainty around what's safe to ship versus what will show up as a visible defect in someone's generated video.
Related Reading
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google to End Free Access to Gemini AI for Non-Subscribers
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
From Archives Uber fulfilled 4 rides from AI agents last quarter out of 320 million
From Archives GPT-6 Astra Generates Game AI Weights Without Training, Beats Trained Models
Common Questions Answered
What is the VBench score difference between full W4A4 quantization and BF16 baseline for Wan2.2-T2V-A14B?
Full W4A4 MXFP4 quantization scored 0.8520 on VBench overall, which represents a 0.0161 gap from the BF16 baseline. Although this numerical difference appears small, AMD's testing revealed that this gap produces visible quality degradation in the generated videos that human reviewers could clearly detect.
Why does AMD recommend W4A8 quantization over full W4A4 for Wan2.2 video generation on AMD Instinct GPUs?
According to AMD's Quark research, both full W4A8 and fixed-rotation/SmoothQuant-selected W4A8 maintain near-BF16 quality levels. The selected W4A8 policy is more practical because it applies 4-bit quantization only where calibration indicates a benefit, rather than uniformly quantizing all weights and activations to 4-bit precision.
What does the 0.0161 VBench gap reveal about pushing quantization limits on Wan2.2?
The 0.0161 VBench gap between full W4A4 and BF16 represents the practical floor for quantization on Wan2.2 before quality visibly degrades. This result serves as a sanity check for users running video generation at scale on AMD Instinct hardware, showing that aggressive full W4A4 quantization sacrifices noticeable video quality despite appearing numerically close to the baseline.
What quantization formats did AMD test for the Wan2.2-T2V-A14B diffusion transformer?
AMD tested multiple quantization approaches including full W4A4 MXFP4, full W4A8, and fixed-rotation/SmoothQuant-selected W4A8 policies on the Wan2.2-T2V-A14B text-to-video generation model. The testing aimed to determine how far quantization could be pushed while maintaining acceptable video quality compared to the full-precision BF16 baseline.
Further Reading
- Technical Dive into AMD MLPerf Inference v6.1 Submission - AMD ROCm Blog
- Wan2.2-T2V-A14B Model Card - AWS Neuron GitHub
- Wan2.2 Fine-Tuning: Tailoring an Advanced Video Generation Model - AMD ROCm Blog
- Wan2.2-T2V-A14B-Diffusers-MXFP8-CT-AutoRound - Hugging Face
- Wan2.2 T2V A14B - Ostris
